Received: 28 August 2017 | Revised: 4 May 2018 | Accepted: 22 July 2018

DOI: 10.1111/jems.12291

OR IG INAL ART I C L E

Can social media lead to labor market discrimination?Evidence from a field experiment

Matthieu Manant* | Serge Pajak | Nicolas Soulié

RITM, Université Paris‐Sud, UniversitéParis‐Saclay, Sceaux, France

CorrespondenceMatthieu Manant, RITM, UniversitéParis‐Sud, Université Paris‐Saclay,Sceaux, France.Email: matthieu.manant@u-psud.fr

Funding informationAgence Nationale de la Recherche, Grant/Award Number: ANR‐10‐CORD‐0002

Abstract

In this paper, we investigate the role of social media as a source of information

for recruiters to discriminate applicants. We set up a field experiment over a

12‐month period, involving more than 800 applications from two fictitious

applicants which differed in their perceived origins, which is an information

available only from their Facebook profiles. During the experiment, an

unexpected change in the Facebook layout reduced the salience of the

information available on social media profiles. Before this change, a significant

41.7% gap between the two applicants callback rates highlights that personal

online profiles are used by recruiters as a source of information to discriminate

against applicants of foreign origin. After the layout change that mitigates our

signal, the difference in callback rates fades away. This result suggests that the

screening conducted by the employers does not go beyond the main pages of

profiles. It also illustrates that design choices made by online platforms may

have important consequences on the extent of discrimination.

1 | INTRODUCTION

Discrimination based on race and ethnic background, or more generally on personal traits, is a major concern. Thespread of social media has made disclosing personal information and traits online a daily habit for hundreds of millionsof people.1 This makes social media attractive to recruiters looking for information on job applicants that does notappear on the application material. While information unrelated to applicants’ professional skills might help improvingthe match between applicants and recruiters, it leaves room for discrimination. In the case of social media used forleisure purposes, for example, Facebook, users often disclose information without considering potential professionalconsequences. Bertrand and Duflo (2017) highlight that the impact of social media on hiring discrimination has beenlargely overlooked in the economic literature so far, although documenting such practices is essential since thediscrimination occurs without the applicants awareness.2 In this paper, we propose an experimental setting to assesswhether information disclosed on social media constitutes a source of discrimination during hiring. More precisely, we

J Econ Manage Strat. 2019;28:225–246. wileyonlinelibrary.com/journal/jems © 2018 Wiley Periodicals, Inc. | 225

*This paper was presented at the Telecom ParisTech Economics and Management Seminar (January 2013), the University Paris‐Est Créteil seminar(January 2013), the Economics and Psychology seminar at Paris School of Economics (March 2013), the 3rd SEEK Conference on Labour Market atZEW Mannheim (April 2013), the 11th IIOC Annual Conference in Boston (May 2013), the Workshop on Experimental Economics in Florence (May2013), the 4th Annual Meeting of the French Experimental Economics Association in Lyon (June 2013), the 62nd Annual Meeting of the FrenchEconomic Association in Aix‐en‐Provence (June 2013), the World Meeting of the Economic Science Association in Zurich (July 2013), the AnnualConference of the European Association of Law and Economics in Warsaw (September 2014), the 29th Annual Conference of the EuropeanEconomic Association in Toulouse (August 2014), the 6th Annual Meeting of the French Experimental Economic Association in Paris (June 2015),the 30th Annual Conference of the European Economic Association in Mannheim (August 2015), the Amsterdam Privacy Conference (October 2015),the LED Université Paris 8 seminar (January 2017), and the Advances in Field Experiments Conference in Chicago (September 2017).

investigate whether recruiters discriminate against job applicants on the basis solely of the information they find onFacebook profiles.

In the literature on labor market discrimination, the negative impact on hiring of personal traits linked tominority groups has been well documented for the sexual orientation, religious beliefs, and political opinions (seeBertrand & Duflo, 2017; Riach & Rich, 2002). In addition, the negative effect of belonging to a racial minority hasbeen widely studied and observed worldwide. Recruiters generally gather information on applicants from theirapplication materials (resume, cover letter, etc.), during the interview, or via word‐of‐mouth (Granovetter, 1995;Rees, 1966). A distinct feature of our study is that the personal traits we use are not disclosed by any of these usualchannels.

The personal traits disclosed by individuals on social media are not revealed for professional purposes but rather toentertain friends and relatives in a broad sense. Several scholars have pointed to the effect of the Internet on labormarkets (Autor, 2001; Kuhn & Mansour, 2014). In relation to the hiring process, Clark and Roberts (2010) reviewdeclarative surveys, which document the emerging use of social media by recruiters to screen applicants. Thosepractices vary widely among the different surveys due most probably to declarative bias stemming from the illegality ofthese practices. It has been shown in other contexts such as the short‐term housing rental market (Edelman, Luca, &Svirsky, 2017), online commerce (Doleac & Stein, 2013), and peer‐to‐peer loans (Pope & Sydnor, 2011) that onlineinformation found by recruiters can be used to discriminate. The paper by Acquisti and Fong (2016) is closest to ourarticle. They use a field experiment to address the issue of discrimination during hiring based on online information inthe US labor market, and find evidence of such discriminatory practices with respect to religious beliefs in some statesand counties. Their project is concurrent but independent from ours.

In this paper, we test the use of social media by recruiters to discriminate in hiring decisions, based on a fieldexperiment on the French labor market. We created two fictitious applicants, which differ in one signal—theirperceived origins—which is available only on their Facebook profiles. The control applicant has a typically Frenchprofile, while the test applicant’s profile reveals that he is from Marrakesh, Morocco (North Africa), and speaksMoroccan Arabic, and so is perceived as a candidate of Arabic origin. People of Arabic descent are a minority in France,and numerous studies on the labor market in France and in other countries show that people of Arabic origin aresubject to hiring discrimination.3 Each applicant was given a unique first name and last name combination on socialmedia to ensure that an Internet search would identify the right profile. Our candidates applied for job positions at anaccountant in the greater Paris area. We sent one application per job opening using pseudo-random assignment method(see Ahmed, Andersson, & Hammarstedt, 2013; Acquisti & Fong, 2016). Sending one application per job openingalleviates risk of detection, and further reduces the burden on recruiters imposed by the experiment. It also allows us touse identical application materials, resumes, and cover letters, for both applicants.4 The pseudo-random assignmentmethod allows us to control, during the experiment, that the two fictitious candidates are applying for similar jobpositions. Following the literature on hiring, we consider a callback from a recruiter to set up a job interview as apositive outcome. Since the two applicants are similar except for their perceived origins displayed only on theirFacebook profiles (hometown and spoken languages), a significant difference in callback rates can stem only from theobservation by the potential employer of this signal, and the use of it to discriminate.

Using our two fictitious applicants, we applied for more than 800 positions over a 12‐month period between March2012 and March 2013, including the main experiment and two robustness checks. For the main experiment betweenMarch and September 2012, we sent 462 applications, equally divided between the two applicants. We find significantlydifferent response rates for our two fictitious applicants: 21.3% for the French candidate, and 13.4% for the Arabicapplicant. This gap in callback rates in favor of the French (vs. the Arabic) applicant suggests that in addition to theinformation in the application material, employers search for other information on applicants, and use the informationfound on Facebook profiles to discriminate. In our case, a small signal on the Facebook profile generates a significantand constant gap of 37% in the probabilities of our two applicants of being called back for interview.5 Thus, personalinformation posted on Facebook has a dramatic effect on the odds of being called for an interview. In other words, weshow that recruiters use information found on Facebook profiles to discriminate among applicants.

We also conduct two robustness checks. First, we test for alternative names, and obtain similar results which showsthat our main result does not rely on the applicants’ names. Second, an exogenous change in December 2012 in thedefault layout of the Facebook profiles allows us to confirm that the gap between our two applicants relies onthe differentiating signal (perceived French origin vs. Arabic origin). The layout change results in a part of our signal,the spoken languages, being located under a secondary tab rather than being displayed on the front page of profiles.After December 2012, the gap in callback rates between the two candidates shrank dramatically, as though in the

226 | MANANT ET AL.

recruiters eyes, both profiles now were identical. The effect of this layout change on the outcome of our experimentconfirms that recruiters were checking our applicants Facebook profiles, and were taking the spoken languages intoaccount when calling back the applicants.

This behavior is in line with the study by Bisin et al. (2008) which highlights the effect of the spoken language as amarker of cultural identity, and its subsequent discriminatory effect for applicants who are perceived as foreigners. Thisresult suggests that an outcome as important as being called back or not for a job interview is influenced largely by the waypersonal information is displayed on social media. The overall experiment confirms the importance of Facebook forrecruiters and gives some insights into the extent of discrimination exercised by recruiters using social media screening.Moreover, the layout change allow us to obtain two additional results. First, the existence of search costs when browsing anentire Facebook profile, which seems to limit the recruiters’ depth of search for personal information. Second, our resultsmake clear that the online platforms design decides how much information is displayed and how salient it is made, whichcan have real consequences in terms of discrimination. Edelman et al. (2017) and Fisman and Luca (2016) both emphasizethis problem when addressing the issues of discrimination in online marketplaces like Airbnb, eBay, or Uber. Hiring isanother area where similar issues arise. Although Facebook is not an online job platform like LinkedIn, it serves animportant role in hiring decision. Our results also have consequences beyond discrimination: Employers are screeninginformation available on personal social media that is not aimed primarily at employers or potential employers. This showsthat the separation between the personal and the professional spheres is becoming increasingly blurred. Also, ourexperimental setting and use of social media, which is in line with Bertrand and Duflo (2017), allows us to test in a naturalcontext for the impact on hiring of extracurricular activities.

The article is organized as follows: Section 2 reviews the labor market discrimination and data‐based discriminationliteratures, with a special focus on digital markets. Section 3 describes the field experiment and our protocol. The mainresults of the experimentation are exposed in Section 4. Robustness checks are provided in Section 5, and Section 6presents some conclusions.

2 | LITERATURE REVIEW

In this paper, we are interested in understanding the use of social media by recruiters, and how it can lead todiscrimination. This investigation contributes to three literature streams. The first refers to the role of nonprofessionalinformation about applicants such as ethnicity, sexual orientation, political opinions, etc., which can lead todiscrimination during the hiring process. The second is related to the use by recruiters of formal and informal channelsof information to gather personal information on applicants or to identify traits. The third strand of work includes anemerging literature on discriminatory outcomes in online markets including the labor market, based on the use ofpersonal information posted online.

2.1 | Personal traits and labor market discrimination

According to the economic literature, hiring discrimination against minority groups relies on two main mechanisms:recruiters dislike of interacting with members of minority groups (Becker, 1957), or the expectation that members ofminority groups are less productive on average than members of other groups (Arrow, 1973; Phelps, 1972). There is alarge empirical literature documenting the different personal traits underlying hiring discrimination, and measuringthe range of the negative effects (Bertrand & Duflo, 2017; Riach & Rich, 1991). These studies rely mostly on fieldexperiments so as to circumvent declarative bias associated with the illegality of this practice. Amongst the personaltraits that can lead to hiring discrimination, the negative effect of belonging to a minority race or ethnic group is themost frequently documented in many contexts (Bertrand & Duflo, 2017).6

Using correspondence test methodology, several studies manipulate applicants’ names or first names so thatapplicants are perceived as members of a racial or an ethnic minority group. In the studies by Bertrand andMullainathan (2004) and Jacquemet and Yannelis (2012) on the American labor market, the names and first names offictitious applicants are manipulated so that they are assigned either typically African American‐sounding names orWhite‐sounding names. Both studies show that individuals with African American names received around half thenumber of callbacks as individuals with White‐sounding names. These results illustrate large and systematicdiscrimination in the United States against the African American minority group compared to the White‐Americanmajority group. Using a similar method of investigation, other studies illustrate the existence of discrimination in hiring

MANANT ET AL. | 227

against racial or ethnic minorities based on applicants’ name in many countries. Oreopoulos (2011) shows hiringdiscrimination in the Canadian labor market against applicants with Chinese, Indian, or Pakistani‐sounding namescompared to applicants with English‐sounding names, that is, applicants from the White majority group. Galarza andYamada (2014) observe discrimination in Peru against individuals with indigenous‐sounding names against individualswith White‐sounding names. In the case of Europe, the results are similar for discrimination against Asian‐soundingnames in Czech and German labor markets (Bartoš, Bauer, Chytilová, & Matějka, 2016). In line with our study, Duguetet al. (2010) and Edo and Jacquemet (2013) provide results of discrimination on the French labor market againstapplicants with Arabic‐sounding names. Alongside these studies that investigate the effect of the names, some studiesaddress the effect of a foreign nationality on labor market discrimination. For instance, Firth (1981) shows that in theBritish labor market, applicants with African, Indian, or Pakistani nationality indicated by their place of birth, havesignificantly lower callback rates than British applicants. In France, Duguet et al. (2010) highlight the negative effect oncallbacks of Moroccan nationality compared to French nationality. Those studies show clearly that applicants that areperceived to be members of a racial or minority group based on name, first name, nationality or cultural characteristicsface hiring discrimination in many countries including France. This evidence from the discrimination literature led usto choose perceived origin (French vs. foreign origin) as a differentiating signal in our experiment. This signal is likelyto generate discrimination. In the empirical literature on labor discrimination, personal traits often are not availableexplicitly but can be inferred from information recruiters find in the application material. This led to our focus on theway recruiters gather applicants’ personal traits, that is, to pinpoint sources of information on applicants.

2.2 | Informal sources of information for recruiters

Formal and informal channels of information can be distinguished based on the information source. In the case ofrecruiters and applications, the formal channel of information is the applicant who provides explicit information torecruiters. This is contained in the application material (resume, cover letter, reference letters, etc.), and, if selected, theinterview. In both cases, the information provided by the applicant as identity, education, work experience, andqualifications can be verified (DeVaro, 2008; Holzer, 1987; Rees, 1966; Spence, 1973).

Informal channels of information for recruiters include the applicants former colleagues and/or acquaintances viareferences and word‐of‐mouth (Albrecht & van Ours, 2006; Granovetter, 1995; Montgomery, 1992; Rees, 1966; VanOmmeren & Wentink, 2012). Since the works by Granovetter (1995) and Rees (1966), many studies have highlighted theimportance of referrals and recommendations in the hiring process (Obukhova & Lan, 2013). Former or current valuedemployees are considered reliable sources of information about potential recruits because they are likely to recommendcandidates with similar competences, and their recommendations can affect their reputations as referees (Rees, 1966;Sterling, 2014). The empirical literature underlines the complex effects of informal contacts and networks on laboroutcomes due to individual, relational, and employer heterogeneity (Cappellari & Tatsiramos, 2015; Ioannides & Loury,2004; Pellizzari, 2010). Studies using more recently available data show that informal channels of information matter sincethey carry reliable information that it would be difficult to convey via formal channels (Bayer, Ross, & Topa, 2008;DeVaro, 2008; Kramarz & Skans, 2014).

The empirical literature on labor market shows that recruiters use both types of information channels (Granovetter, 1995;Rees, 1966). However, as Autor (2001) and Kuhn (2014) suggest, labor market and hiring process are being affected stronglyby the Internet. Some studies, based mainly on declarative surveys, to the use of the Internet, and especially social media asemerging channels of information on applicants (Clark & Roberts, 2010). However, very few works have intended to focuson the potential discriminatory behaviors that can originate from this new practice.

2.3 | Discriminatory outcomes on online markets

Our article also contributes to the growing number of studies focusing on use of online information todiscriminate in online markets (Fisman & Luca, 2016). In online selling of used goods, Doleac and Stein (2013)show that if a picture attached to an ad conveys information about the sellers race, this has consequences for thenegotiation with prospective buyers. African American sellers receive 17% fewer offers and a smaller highestoffer compared to their white counterparts. In an online peer‐to‐peer lending market, Pope and Sydnor (2011)show that, for similar credit profiles, White (vs. African American) borrowers enjoy both a higher chance ofbeing funded and a lower subsequent interest rate. In the short‐term rental housing market, Edelman et al.(2017) create fictitious guests with distinct characteristics and are able to document an eight‐point penalty in the

228 | MANANT ET AL.

acceptance rate for guests with typically African American names. Beyond the human decision to discriminate,Lambrecht and Tucker (2018) present evidence of the impact of the auction algorithm which is used to allocateadvertising impressions and its discriminatory outcome.

In a labor‐market context, Acquisti and Fong (2016) is the closest to this paper. They address the question ofdiscrimination by recruiters in the United States based on personal traits related to religious belief (Christian vs.Muslim) and sexual orientation (straight vs. homosexual) available on social media.7 They implement acorrespondence test using two pairs of fictitious applicants, and manipulated a signal only available on Facebookmade with closed‐ or open‐ended text fields: “Religion:” (Christian or Muslim), and “I’m interested in:” (femaleor male). They submitted around thousand applications in various American cities for each fictitious applicant,and found a significant difference in callback rates only for religious treatment but not at the national level.These gaps in callback rates occur in states and counties characterized by a high proportion of Republican voters,relatively older people, and a low proportion of Muslims. Their setting and results differ from ours in severalways. First, we are only interested on the impact of an applicant’s perceived foreign origin. This personal traithas been shown by Duguet et al. (2010) and Edo and Jacquemet (2013) to induce strong name‐baseddiscrimination in the same geographical area as our study (the greater Paris area). Second, our indication offoreign origin is based on text only, while Acquisti and Fong (2016) also manipulate the background image acrossfictitious applicants. Third, their applicants have a large online presence, for example, LinkedIn, Google+, andGoogle Sites. This wider presence tends to provide several sources of online information for recruiters, which canmitigate the effect of the manipulation. Fourth, during the course of our experiment, the layout of the Facebookprofiles changed and displaced some information into a subtab. Acquisti and Fong (2016)’s study was notaffected by this layout change since it occurred before the beginning of their experimentation. As a consequenceof this natural experiment, we are able to compare two versions of the site’s visual ergonomy. This allows us toget a sense of how deep into the profiles recruiters were looking for information, and to determine that thescreening is very superficial. It illustrates that small design choices made by the platform in how prominently itdisplays user information can have major consequences.

3 | EXPERIMENTAL DESIGN

We conduct a correspondence test using pseudo-random assignment to detect whether employers discriminate basedon information they found on social media. In this section, we describe the design of our experiment, that is, thecreation of two applicants with identical resumes and cover letters, and with identical social media profiles except forthe tested signal. Below we describe our methodological choices (3.1), the characteristics of the fictitious applicants(3.2), and our applications protocol (3.3).

3.1 | Methodology

3.1.1 | Labor market and field experiments

To capture whether employers’ real practices rely on information found on social media and to avoid declarative bias,we chose a field experiment. In the labor market literature, field experiments rely on two main methodologies: audit/situation testing, and correspondence testing. Audit testing consists of real people, usually professional actors briefed bythe experimenters, who apply for job openings, and present themselves for job interviews. The audit approach allows afocus on employers’ hiring behaviors along the multiple steps of the hiring process, that is, (a) whether they call thecandidate for an interview, (b) whether they offer the position after the interview, and (c) the wage level offered. One ofthe limitations of this approach is that it is difficult in practice, to ensure similar applicant performance in face‐to‐faceinterviews. Also, actors’ direct interactions with employers during interviews raise concerns about experimenter bias.8

Correspondence testing involves fewer methodological issues (Riach & Rich, 2002). It allows for better control over theexperimental environment, especially the content of applications. The method is also less time‐consuming and easier toreproduce (Bursell, 2007). In addition, although correspondence testing reveals discrimination at the interview callbackstage, and not at the final hiring decision stage, it has been shown that about 90% of the discrimination occurs in theformer stage (Riach & Rich, 2002). According to Bursell (2007), correspondence testing is a type of randomizedexperiment, and therefore, provides the most convincing method to allow causal inferences.9 The main difficulty lies inconstructing two applications that are similar in all the relevant characteristics except the one being tested. Indeed, the

MANANT ET AL. | 229

two applicant materials need to be sufficiently different to avoid detection by the employer, but sufficiently similar thatthe difference in callback can be attributed to the tested characteristic.

3.1.2 | Pseudo-random assignment

Following Ahmed et al. (2013), we carry out a correspondence test in which we send only one application per job opening.Compared to the usual correspondence testing which consists of sending at least two applications per job opening, thismethodology has three main advantages. First, it allows us to create two fictitious applicants that are identical regarding theirapplication material (resumes and cover letters) and with identical Facebook profiles apart from the manipulated characteristic(or signal) information that is available only on their social media profiles. Any observed difference in the callback rates maythen be clearly attributable to this signal. Second, since the applications are sent to two different firm samples, there is no risk ofdetection by these firms. Third, each recruiter receives only one application. This means that our study interferes onlymarginally with the real hiring process compared to sending multiple applications for each opening. This is more efficient andmore desirable ethically.10 We ensure that our two applicants apply for positions with similar characteristics using a pseudo-random assignment procedure by responding to job openings available on pole‐emploi.fr, the website of the French publicemployment agency Pôle Emploi. Ads on this website provide detailed information on both job positions and firms.

In line with the literature, we consider a callback from a recruiter to arrange a job interview as a positive outcome.Since the two applicants are similar except for one signal that is available only on their Facebook profiles, a significantdifference in callback rates may stem from observation of the signal by recruiters. Another approach would be to use asystem similar to Google Analytics to track whether recruiters visited our candidates’ profiles. However, the datagathered would be virtually useless for economic analysis. For example, if the information relied on the recruiters’ IPaddress, it would be impossible to distinguish the companies’ address from the Internet service providers IP address,and particularly in the case of smaller firms.11 The most meaningful statistic produced by an analytics tool would be thetotal number of visits to the profile, and not the part of the recruiters who visited it. Finally, it is not possible to includean analytics code in the Facebook profiles.

3.2 | Fictitious applicants

3.2.1 | Resumes and cover letters

The construction of the application material has three objectives: (a) creation of similar material for both applicants, (b)producing realistic material, and (c) maximizing the number of interview calls so as to increase statistical significance. Sinceour two applicants have the same resumes and cover letters, the two sets of application materials are similar. Before startingthe experiment and alongside conducting pretests, we interviewed human resources managers to ensure the relevance of thecover letters to current job market conditions. Also, depending on what the job ad specified, both the letter and the resumeincluded some standard sentences corresponding to profiles being sought by the recruiters.12 Cover letters and resumes thatare too general and too standardized are usually considered less effective because most employers are seeking a specificprofessional profile (customer, supplier, asset manager, etc.). We used information on the recruiting firm available on theInternet on official websites, web articles, etc. When this information on the firms was not available, the cover letters wereunspecific. Resumes and cover letters were submitted by e‐mail as pdf files accompanied by a standard message. Tomaximize hiring opportunities, both our applicants declared being already in employment when they applied, and havingnever been unemployed.13 Both have a 3‐year higher education degree in accounting, flawless school record, and threeinternships with various experience suitable for most accounting jobs. We decided to focus on accounting jobs for two mainreasons. First, accountancy is a back‐office job, which usually involves no direct contact with customers. This avoidsdiscrimination based on customer preferences that might affect the hiring process. Secondly, accountant positions are amongthe most regularly advertised jobs on the Pôle Emploi website in many different sectors, and thus, ensured a minimumnumber of ads available to apply to each week. It avoided a focus on a few sectors with specific hiring practices.

For each applicant, in both the resume and the cover letter, we provide the following information: name, address,owning a driving license, date of birth and age, phone number, and e‐mail address. In the applications package, there wasno link to the applicants’ social media, nor any reference to the personal traits manipulated. The applicant’s address is inan affluent district of Paris (15th arrondissement) which avoids location‐based discrimination. The applicant holds adriving license. The phone numbers are distinct in each resume so as to track candidates’ callbacks. The e‐mail address ofeach applicant is registered on Gmail with the user names following the same pattern “firstname.lastname@gmail.com.”

230 | MANANT ET AL.

3.2.2 | Social media profiles

In our experiment, the only difference between our two fictitious applicants relies on a signal appearing only on theirFacebook profiles. The social media profile of one applicant includes a signal that could lead him to be perceived as amember of the Arabic minority. More precisely, the profile mentions that the applicant is from a well‐known North‐Africancity (Marrakesh, Morocco), and that the applicant speaks Moroccan Arabic. The control profile refers to a typical Frenchcity (Brive‐la‐Gaillarde, France), and the applicant speaks Italian. Thus, we differentiate our applicants by city of origin andlanguages spoken according to their social media profiles (see Figure 1 for screenshots of the two Facebook profiles).

We chose this signal because many studies highlight the negative impact of an applicant’s perceived origins or ethnicityon interview call rates (Ahmed & Hammarstedt, 2008; Bertrand & Mullainathan, 2004; Duguet et al., 2010; Jacquemet &Yannelis, 2012). In particular, studies conducted in Europe show that applicants’ foreign language skills act as a signal offoreign cultural identity (Battu & Zenou, 2010; Bisin et al., 2008). Thus, this signal can produce discrimination if thesupposed cultural identity is different from the recruiter’s, or might reinforce the perceived cultural similarities between theapplicant and the recruiter (Edo & Jacquemet, 2013). We find the studies by Duguet et al. (2010) and Edo and Jacquemet(2013) particularly interesting since we use very similar fictitious applicants in terms of education, qualifications, and jobsearch area (i.e., Paris region). Both these studies used correspondence tests to measure the extent of discrimination againstArabic applicants for accountant or assistant‐accountant positions in the Paris region. They show a gap of 35–40% in callbackrates in favor of the French compared to the Arabic applicants. Therefore, provided that the signal is perceived, we expect itto have a negative impact on the odds of being called for interview.

To ensure that the foreign origin signal is conveyed only by the social media profile, the first and last names of eachapplicant are French‐sounding, namely Thomas Marvaux and Stéphane Marcueil. The first names were picked from the top‐five first names for the year of birth. Each first name‐last name combination is unique on Facebook,14 and our fictitiousprofiles are the sole results of searches on these first and last names on the three leading French language web searchengines, and on Facebook.15

3.3 | Experiment protocol

We selected job openings published between March 19, 2012 and September 30, 2012 on the French public agency foremployment website—Pôle Emploi. We focus on this website because job openings advertised on this website providesystematic and detailed information on jobs (wage, type of contract, working hours, education requirements, workexperience, etc.), and firms (name, location, sector, size, etc.). This information is crucial for pseudo-random attributionof applicants to job offers and for statistical analysis. Other popular French employment websites—Monster, Keljob,Indeed, etc.—provide fewer details, and therefore, were not considered. Moreover, we selected only openings thatprovided a recruiter’s direct contact information (contact name and e‐mail), and excluded those which required theapplicant to contact a third party, usually Pôle Emploi or a recruitment agency. Only openings for long‐term workrelations were considered, that is, with regular (undefined work duration) or fixed‐term contracts (at least 6 months). InFrance, ending these types of contracts involves high severance payments in addition to legal and administrativeprocedures.16 In this context, recruiters are expected to be more careful when screening applicants.

Our applicants have a bachelor’s degree (i.e., 3 years of undergraduate education) in accounting, and we respondedto ads in the three relevant groups (accountant, assistant‐accountant, and aid‐accountant) in the Pôle Emploicategorization. For each selected opening, we generated application materials, that is, the resume and the cover letter,using predefined key sentences to match the advertised job. To avoid experimenter bias, the material was not assignedto an applicant until all the application material had been generated.

We observe the recruiters’ behavior based on the differences (if any) in return rates for the two fictitious applicants.The candidates have identical application packages and differ only in selected information available on their socialmedia profiles. Following Acquisti and Fong (2016) and Ahmed et al. (2013), we sent only one application per job offerand use a pseudo-random assignment procedure to ensure that both applicants apply to similar positions along theexperiment. This pseudo-random assignment is based on job positions (accountant, assistant‐accountant, etc.), requiredwork experiences, firm sizes, and firm sectors (see Table 1 for descriptive statistics). Half of the recruiters in our samplereceived an application from the French candidate, and half received an application from the Arabic candidate. Figure 2summarizes the timing of the experiment.

MANANT ET AL. | 231

4 | RESULTS

The results of the experiment are as follows. First, we present the main outcome of the experiment based on graphicsand independence tests (model‐free evidence). Second, we provide robustness checks of the results using probit models.In Section 5, we conduct two additional robustness checks related to the names of the fictitious applicants, and thesignal perceived by recruiters.

FIGURE 1 Screenshots of the social media profiles of our two applicants and the differentiating signal (city of origin and spokenlanguage) [Color figure can be viewed at wileyonlinelibrary.com]

FIGURE 2 Experiment timing

232 | MANANT ET AL.

4.1 | Model‐free evidence

We sent a total of 462 applications from both candidates between March 19, 2012 and September 30, 2012. Applicationswere sent at similar rates over this period. Each job opening received one application, and we ensured that the twoapplicants applied for positions that were similar according to the observable characteristics of job positions, usingpseudo-random attribution process. Table 1 presents a description of the job positions for each applicant.

Table 2 below provides an overview of the main result of this paper. During the period, our applicants received 80positive calls for interview. This overall return rate is quite high (17.3%) compared to other studies on the Paris area(Duguet et al., 2005, 2010; Edo & Jacquemet, 2013). The most plausible explanation for this higher interview call rate inour study is that we matched application cover letter and resume to each job opening, incorporating predefinedsentences corresponding to the characteristics of the firm and the advertised position. Recruiters contacted ourapplicants mostly by phone (68.0%) and e‐mail (26.3%). A few used both means (5.7%). The applicants’ postal addresseswere real but did not receive regular mail.

TABLE 1 Description of the job applications for each applicant, March–September 2012

Variables French applicant Arabic applicant

Job position (%)Accountant 44.3 43.1Specialized accountant 9.6 10.3Accounting assistant 23.9 29.3Accountant and secretary assistant 18.7 14.7Other accounting assistant 3.5 2.6

Required education (%)Not specified 40.4 38.8Vocational certificate 1.3 2.1High school diploma 11.3a 5.2a

Associate degree 42.6 49.6Bachelor’s degree 4.4 4.3

Required work experience (%)No experience 25.2 27.16 months–1 year 12.6 12.52 years 27.8 23.33 years 16.5 19.44 or 5 years 17.9 17.7

Firm size (%)0–5 employees 23.5 23.36–19 employees 27.8 29.720–49 employees 17.8 19.050–249 employees 21.3 21.1250+ employees 9.6 6.9

Firm status (%)Private 83.0 83.6Public 6.5 4.7Not‐for‐profit 10.4 11.6

Contract (%)Regular 28.7a 19.4a

Fixed‐term 71.3a 80.6a

Fixed‐term contract length (mean [sd], month) 8.2 (4.1) 7.8 (3.3)

Work timeb (mean [sd], hr/week) 34.3 (5.4) 34.2 (5.6)

Mean wage (mean [sd], €/hr) 12.3 (2.4) 12.3 (2.2)

Industry see Appendix A

Location see Appendix A

Number of applications 230 232aIndicates a significant difference in proportion at 5% threshold between the two applicants.bFrench standard weekly work time = 35 hr.

MANANT ET AL. | 233

The number of interview calls received by our two applicants shows a consistent gap in favor of the Frenchapplicant. In particular, Figure 3 shows a roughly constant gap of 8% points between the two applicants in favor of thecontrol candidate who received 21.3% positive returns compared to 13.4% for the test candidate. A χ2 test ( χ2(1) = 5.09;Pr = 0.024) and Fisher’s exact test (Pr = 0.027) confirms the significance of this difference. This gap represents a drop of37% in callbacks for the Arabic candidate. Figure 3 shows also that both candidates experienced an increased averagenumber of callbacks after June 2012. This change can be explained by the easing of the labor market after the Frenchpresidential elections in April–May of 2012. The number of applications sent every month follows the same evolution,showing that firms had more confidence in the economic context after the elections.

Based on the experimental protocol in place, this gap in favor of the control candidate must be the joint result ofemployers choosing (a) to screen applicants’ Facebook profiles, and (b) to use the information collected to discriminatecandidates. Employers use information obtained from social media, and consider this information as reliable although it isnot part of the formal application package. The results show that the content of the online profile is taken into account forthe decision to interview the candidate. In other words, personal information displayed on social media profiles has becomea part of the application material considered by recruiters, and can be used to discriminate applicants.

4.2 | Main results

To confirm our first result, we use a probit model to control for different variables that might affect the probability ofbeing invited for interview. For an application i, the probability of being called back is modeled as follows:

Pr Interview β ArabicApplicant β ApplicationDelay β JobCharacteristics β FirmCharacteristics

β Time α

( ) = + + + +

+ + + ϵ ,

i i i i i

i i

1 2 3 4

5 (1)

where ArabicApplicanti is a dummy variable that equals 1 for the Arabic applicant, and zero for the French one.ApplicantDelayi is a set of dummy variables indicating the number of days between the posting of the job opening onthe website and the application. JobCharacteristicsi includes various job characteristics such as contract type, jobposition, mean wage offer, etc., related to the job offer.17 FirmCharacteristicsi is a vector of characteristics related to the

TABLE 2 Callback statistics by applicant

French applicant Arabic applicant Total

Positive callbacks 49 31 8021.3% 13.4% 17.3%

Negative callbacks 181 201 38278.7% 86.6% 82.7%

Total 230 232 462

FIGURE 3 Applicants callback rates [Color figure can be viewed at wileyonlinelibrary.com]

234 | MANANT ET AL.

firm that offers the job such as industry, size, location, etc. Timei is a set of monthly dummies, while α and ϵi are,respectively, a constant and an error term.

Table 3 provides three specifications of the probit model described above (see Appendix 1 in Supplemental materialsfor detailed results). Model 1 contains only the main explanatory variable, plus time dummies and a constant. Model 2includes all the available control variables, and Model 3 includes all the control variables except “mean wage offer” dueto multicollinearity issues.18 In all three specifications, the Arabic signal that is available only on the online profile isassociated with a significant negative impact on the probability of being invited for an interview. These results confirmthat social media is a source of information that is being used by recruiters to discriminate applicants without theirawareness.

5 | ROBUSTNESS CHECKS

Here, we propose two robustness checks on our results. First, we test whether changing the applicants’ names has animpact on the outcome of the experiment. Second, we take advantage of a Facebook webpage redesign that occurred inmid‐December, namely a change in the profile layout, to provide evidence of recruiters’ perception of informationdisplayed on the social media.

5.1 | Alternative names

The design of our experiment is such that the only difference between our two applicants is the Arabic signal on theFacebook profile. However, the names and first names of the applicants, although they sound French and are similar, canhave an effect on the application outcome. To control for this effect, from October 2012 to March 2013, we used a second pairof fictitious applicants with new names and first names. Again, the new first names were selected from among the five mostpopular French‐sounding first names, and again, the two French‐sounding last names are very similar. These combinationsalso are unique on Facebook. We reset the new fictitious applicants work experience so that our second pair of applicantsgraduated in July 2012. Thus, the new candidates have less work experience ranging from 1 month to 6.5 months. Thisavoided our fictitious applicants having had significantly more work experience at the end of the experiment; discriminationagainst foreign origin individuals is known to decrease with applicants’ work experience (Aeberhardt, Coudin, & Rathelot,2010). Table 4 summarizes the names of the applicants and their work experience in the overall experiment, and Figure 4presents the timing of the experiment with the second set of names.19

Figure 5 shows that applicants of the same type but with different names, one month after applying for job positionshave a similar cumulative callback rate. Also, this callback rate is higher for the French applicants compared to theArabic applicants.20 This comparison was applied only to the first two months (60 days) for each pair of applicantsbecause the Facebook layout change occurred 2 months after we introduced the second pair.

TABLE 3 Probit model results

Dependent variable: callback (yes/no) Model 1 Model 2 Model 3

Arabic applicant −0.313** −0.431*** −0.404***

(0.139) (0.159) (0.156)

Application delay controls No Yes Yes

Job characteristic controls No Yes Yes

Mean wage offer No Yes No

Firms’ characteristic controls No Yes Yes

Time dummies Yes Yes Yes

Constant −0.567** 1.826** 0.386

(0.234) (0.784) (0.592)

Observations 462 462 462

Pseudo‐R2 0.051 0.232 0.216

Note. Robust standard errors in parentheses. *, **, and *** mean significant at 10%, 5%, and 1% thresholds. Omitted variable for applicant type is Frenchapplicant.

MANANT ET AL. | 235

Independence tests confirm that the several last names and first names used for our fictitious applicants have no impacton the experiment outcome. The callback rates of the Arabic applicants are not different (χ2(1) = 0.46;Pr = 0.496, two‐sided Fisher’exact test (Pr= 0.532) and one‐sided Fisher’ exact test (Pr= 0.363)). This is also the case forthe French applicants (χ2(1) = 0.005; Pr = 0.944, two‐sided Fisher’exact test (Pr= 0.99) and one‐sided Fisher’exact test (Pr =0.561)). Additional evidence that applicants’ names do not affect the outcome of the experiment is presented in the next section.

5.2 | Social media layout change

In December 2012, two months after we had been using alternatives names, the social media platform rolled‐out a newdefault layout for the profiles, so that our signal was split into two smaller signals on different tabs.21 The profile layoutchanged from a single page, to a front page with tabs to provide access to certain personal information. Specifically,information on city of origin continued to be displayed on the front page (default “Timeline” tab requiring no click from theviewer) but information on language(s) required a click on the “About” tab and a scroll down the page to be accessed.Figures 6 and 7 below are screenshots of Facebook’s new layout. As the default layout was rolled‐out in mid‐December, forthe positive callbacks received in this timeframe it is not possible to determine whether the recruiters based their decisionson the new layout or not. To suppress any uncertainty during the transition period, we consider as the period before thelayout change as the applications sent before December and which received their positive callbacks before December.

This exogenous change allows us to better understand the way the signal is perceived by recruiters. This modificationcould have created a search cost to access the entire signal (i.e., the city of origin and the languages spoken). If themechanism we suggest holds, that is, that recruiters do look at the Facebook profiles, then the effect should become less

TABLE 4 Applicants’ names and first names used in the experiment

Timespan First name and name Type of applicant Work experience

March 2012–September 2012 Thomas Marvaux French 6.5 to 13 monthsStéphane Marcueil Moroccan

October 2012–March 2013 Julien Bautrant French 1 to 6.5 monthsNicolas Lautrant Moroccan

FIGURE 4 Experiment timing with the alternative names

FIGURE 5 Callback rates in the first two months for each applicant [Color figure can be viewed at wileyonlinelibrary.com]

236 | MANANT ET AL.

significant as the information is harder to retrieve. This intuition is confirmed since the layout change greatly affectedperception of the differentiating signal by recruiters; the constant eight‐point gap we observed since the beginning of theexperiment no longer held after December 2012. This result appears to be coherent with the literature that highlights therole of spoken languages as a significant factor in job discrimination (Bisin et al., 2008; Edo & Jacquemet, 2013;Oreopoulos, 2009). We present detailed results below.

Table 5 and Figure 8 below show clearly that callback rates before and after the layout change are different. Beforethe layout change (October–November 2012), Table 5 shows that the Arabic applicant received four positive callbacksfrom a total of 69 applications, and the French applicant received 13 positive callbacks from a total of 81 applications,that is, a 10% point gap in callback rates between the French (16.0%) and the Arabic (5.8%) candidates. A χ2 testindicates the significance of this difference at the 5% threshold (χ2(1) = 3.90; Pr = 0.048).22 This result is additionalevidence that our previous findings do not arise from our applicants’ names and first names. The period before thelayout change (March–November 2012) confirms our main result. Indeed, in the first 8 months of the experiment withthe initial Facebook layout, the French applicant received 62 positive callbacks from 311 applicants (19.9%) compared to35 callbacks from 301 applications (11.6%) for the Arabic applicant (χ2(1) = 7.91; Pr = 0.005).23 This corresponds to a41.7% decline which is in line with the French studies testing for name‐based discrimination mentioned earlier whichfind a 35–40% decline for Arabic‐sounding applicants (Duguet et al., 2010; Edo & Jacquemet, 2013). In those studieshowever, the information signaling the applicant’s foreign origin is directly available in their resumes. Our settingintroduces the information only in the Facebook profiles, but leads to a similar impact in terms of discrimination. Theshare of recruiters browsing the applicants profiles is not a data made available by Facebook. However,the discrimination rate in our experiment is similar to that in studies based on the applicants names – as if theonline information that we introduce is perfectly obtained by employers. Assuming that recruiters discriminatorypractices are consistent with the studies of Duguet et al. (2010) and Edo & Jacquemet (2013), the gap between callbackrates seems to indicate that most employers searched for the information, and obtained it from social media.

After the Facebook layout change (i.e., after December 20th, 2012), the Arabic applicant is called back at a rate thatis not statistically different from that of the French applicant, respectively 13.2% and 9.5% (χ2(1) = 0.64; Pr = 0.42).24 To

FIGURE 6 Screenshot of the social media profile of the Arabic applicant with the new layout: Default “Timeline” tab displaying onlythe town of origin (same kind of display for the French applicant) [Color figure can be viewed at wileyonlinelibrary.com]

MANANT ET AL. | 237

better understand the consequences of the layout change, we present in the table below the respective effects of the newlayout and of being an Arabic applicant on the probability of receiving a callback. More precisely, we estimate thefollowing equation for an application i using a probit model:

Pr Interview β ArabicApplicant LayoutChange β ApplicationDelay β JobCharacteristics

β FirmCharacteristics β Time α

( ) = + × + +

+ + + + ϵ ,

i i i i i

i i i

1 2 3

4 5 (2)

where LayoutChangei is a dummy variable equals to 1 after the social media layout change (i.e. after December 20,2012), and zero before the December 1, 2012. The other variable or set of variables used are similar to those described inthe “Main results” section. The introduction in the model of an interaction term, which indicates the Facebook layoutchange, allows us to strengthen and deepen our previous results.

FIGURE 7 Screenshot of the social media profile of the Arabic applicant with the new layout: “About” tab displaying the town of originand the spoken languages (same kind of display for the French applicant) [Color figure can be viewed at wileyonlinelibrary.com]

TABLE 5 Callback rates before and after the Facebook layout change with alternative names

Before layout change After layout change Overall experimenta

October–November 2012 December 20, 2012–March 2013 October 2012–March 2013

French app. Arabic app. French app. Arabic app. Total

Positive callbacks 13 4 9 12 3816.0% 5.8% 9.5% 13.2% 12.8%

Negative callbacks 68 65 86 79 29884.0% 92.9% 90.5% 86.8% 87.2%

Total 81 69 95 91 336aThe overall experiment period excludes the period of transition to the new layout, December 1–19, 2012.

238 | MANANT ET AL.

For the second pair of fictitious applicants, Table 6 shows that before the layout change we still observe a negativeand significant effect of the Arabic candidate compared to the French candidate (See Appendix 6 in SupportingInformation materials for detailed results). This result confirms that Facebook has become a new informal source ofinformation for recruiters which enables early‐stage discrimination.

The second interesting point is that there is no longer a difference between the French applicant with the old layoutand the Arabic applicant with the new layout (i.e., without the spoken language appearing on the front page). In otherwords, once the differentiating signal is mitigated (removal of spoken languages from the front page) the two applicantsare no longer considered different by recruiters. The similarity of the two fictitious applicants becomes more noticeableand the role of spoken language as the main differentiating factor is highlighted. This result is in line with other studieson discrimination showing that languages are important explanatory factors in job discrimination (Bisin et al., 2008;Edo, Jacquemet, & Yannelis, 2015; Oreopoulos, 2009). The second element of our signal, namely “city of origin,” doesnot seem to have an impact if isolated from the spoken languages. Bisin et al. (2008) show that foreign spoken languageseems to act as a signal of origin in a foreign culture.

FIGURE 8 Monthly callback rates [Color figure can be viewed at wileyonlinelibrary.com]

TABLE 6 Layout change impact on application outcome

Dependent variable: callback (yes/no)

French applicant and new layout −0.736***

(0.289)

Arabic applicant and old layout −1.122***

(0.339)

Arabic applicant and new layout −0.340

(0.277)

Application delay controls Yes

Job characteristic controls Yes

Mean wage offer control No

Firm characteristic controls Yes

Time dummies No

Constant 1.438***

(0.805)

Observations 336

Pseudo‐R2 0.209

Note. Robust standard errors in parentheses. *, **, and *** mean respectively significant at 10%, 5% and 1% thresholds. Omitted variable for applicant type andlayout is French applicant and old layout. In this model, layout change dummy acts as time dummies.

MANANT ET AL. | 239

This result points also to the existence of search costs associated with screening social media platforms. Exploring allthe tabs seems to be costly; recruiters tend not to carry out a thorough screening of applicants’ personal social mediaprofiles, and to rely only on the front page. More study is required to characterize this behavior further. For instance, inour experiment the education level of applicants is 3 years undergraduate study; it is possible that more senior jobs wouldinvolve more thorough social media profile screening. Moreover, it shows that online platform design matters when itcomes to the extent of discrimination (Edelman et al., 2017; Fisman & Luca, 2016). A small change in the platform design,by making an information still available but less salient, dramatically lowers the resulting discrimination.

6 | CONCLUSION

Advancements in social media use means that multiple personal traits can be identified by screening job applicantsonline. Our findings show that the consequences in the hiring process are real: the audience of leisure‐oriented socialmedia includes professionals. For potential employers inclined to discriminate against applicants, a candidate’s socialmedia profile carries considerable weight. These results are obtained by setting an experiment using real jobapplications for accountant positions in the greater Paris area. We create two fictitious applicants who differed only intheir perceived origins, observable only on their Facebook profiles. While the control applicant is associated to atypically French hometown and to European spoken languages, the test applicant displays signals that means that he isperceived as member of the Arabic ethnic minority in France. His Facebook profile indicates Marrakesh (Morocco,North Africa) as his city of origin, and Moroccan Arabic as a spoken language. In line with the literature on labordiscrimination, this signal—if observed—is expected to have a significant negative effect on rates of callback forinterview for the test applicant compared to the control applicant (Bertrand & Duflo, 2017). We send more than 800applications for job openings for accountants in the Paris region using the pseudo-random assignment method. Duringthe experiment, an unexpected change in the Facebook layout altered the display of our online signal by sending awaythe applicant’s language spoken into a subtab. We take advantage from this natural experiment.

Before the layout change, in the period March to November 2012, among more than 600 applications callbacksshow a constant and significant gap of 8% points between the French (19.9%) and the Arabic (11.6%) applicants.The design of the experiment and the subsequent robustness checks ensure that the 41.7% difference in callbackrates resulted solely from observation of the Arabic signal available only on the Facebook profile. This significantdifference indicates that personal social media are a new informal source of information on applicants that canlead to discriminatory callback decisions. In implementing a field experiment to shed light on these newdiscriminatory practices our paper adds to Acquisiti and Fong (2016) work which is concurrent but independentfrom our experimentation. For many recruiters, Facebook profiles are being considered a part of the applicationmaterial along with the applicants resume and cover letter, and social media are affecting the matching process inthe labor market. After the layout change, discrimination fades away with callback rates no longer statisticallydifferent. Our results illustrate the existence of search costs associated with online screening, and give clearevidence that online platforms design choices in terms of information salience may influence the extent ofdiscrimination, as emphasized by Edelman et al. (2017) and Fisman & Luca (2016).

We identify some limitations regarding the generalization of our results.25 Our experimental setting is realistic butcomes with some necessary simplifications. For instance, the content of our Facebook profiles was basic but realistic.An actual Facebook profile can include more information about its owner, and as a result, send a mixed signal to therecruiter. In our setting, we simplified the complexity of the Facebook profiles to allow for a clear interpretation of ourresults. Similarly, choosing a unique name and first name combination for applicants ensures that the recruiter does notfind anyone other than our fictitious candidate. Actual applicants can have homonyms which would impede findingtheir profile and would reduce the gap in callbacks between the two profiles. However, this would be a complex settingwhere the final effect could not be disentangled. It also would raise an ethical issue as the fictitious candidate couldhave unwanted effects on the existing homonyms. Finally, we are measuring callbacks not actual hiring decisions.Although it is not equivalent to the final decision to hire, it shows a lower employment opportunity: we are testing theearly stage of hiring which gives a lower bound of discriminatory screening since the actual hiring has additional stepswhere discrimination can still occur.

Our findings have three implications for public policy. The first is related to job applicants who shouldunderstand that their social media presence is part of their application material. Government communicationsshould include this aspect when educating the public. Our findings suggest potential solutions to its use, ranging

240 | MANANT ET AL.

from locking one’s profile, to—where possible—cleaning information that might lead to discrimination during jobsearch or assessment periods, to use of multiple social media profiles (“official” profile accessible to anyone, andprivate), to use of an avatar. The second implication is related to social media platforms. The second part of ourexperiment makes clear that a change in the layout of the social profile can have a dramatic effect on theinformation gathered by the employer, and consequently can affect the employers decision. The social profilelayout acts as what Sunstein and Thaler (2008) describe as a nudge, and drives the decision by framing theinformation presented to the audience. The public debate over social media has so far, largely evaded theirresponsibility in terms of which information is made salient and the consequences of this design choice. Onlineplatforms should be aware of the potential consequences their design choices. For instance, Fisman and Luca(2016) recommends that sensitive information such as gender or race be made either not accessible or less salient.Third, policymakers and other organizations tackling discrimination should be cognizant that the source ofdiscrimination is not just the formal application material which is already well regulated and tested. In line withBertrand and Duflo (2017), our results suggest that a particular focus should be put on extracurricular information(“beyond the resume”) available online.

ACKNOWLEDGMENTS

We thank the audiences for their comments and feedback. We are grateful to Alessandro Acquisti, Anja Lambrecht,Catherine Tucker, and Pai‐Ling Yin for their insightful comments on this study. We thank Alain Rallet, FabriceRochelandet, Fabrice Le Guel, and Grazia Cecere for their comments and support during the experimentation. Thisstudy was supported by grant ANR‐10‐CORD‐0002 from the French National Agency for Research. The University ofParis‐Sud does not have an IRB for nonmedical research.

END NOTES

1 For instance, Facebook, the most popular social media, had reached more than 2 billion monthly active users in mid 2017. Information about Facebook users(birthdate, occupation, etc.) is often publicly available, even to non-users. Access to more detailed information requires only the completion of a freeregistration form.

2 For most social media, and especially Facebook, it is impossible to know whether someone has screened a profile unless the individual leaves an explicit trace(a message, a like, etc.). For recruiters, this feature of social media sites is important since use of personal information—whether found online or elsewhere—to select applicants is illegal in many regions.

3 The literature on labor market discrimination shows that this characteristic negatively affects an applicant’s callback rate and labor market integration moregenerally. See Duguet and Petit (2005); Duguet, Leandri, L’Horty, and Petit (2010); Edo and Jacquemet (2013) for France; Bertrand and Mullainathan (2004)for the United States; Battu and Zenou (2010); Bisin, Patacchini, Verdier, and Zenou (2008) for the UK.

4 In the usual systematic attribution testing methodology, two or more applications for a job opening are sent simultaneously, and the resumes and cover lettersof the fictitious applicants need to be sufficiently different to avoid detection by an employer.

5 This gap reaches 41.7% when including applications sent until the layout change.6 Other personal traits leading to labor market discrimination include sexual orientation (Ahmed et al., 2013; Drydakis, 2009; Weichselbaumer, 2003), gender(Booth & Leigh, 2010) and attractiveness (Galarza & Yamada, 2014).

7 Our field experiment is independent from Acquisti and Fongs (2016) study, which started in early 2013. Ours was conducted from early 2012 to early 2013.8 See Bertrand & Mullainathan (2004); Bursell (2007) for other labor market‐related audits results and limitations.9 For examples of correspondence tests, see Brown and Gay (1985) and Hubbuck, Britain, and Carter (1980) for the UK, Riach and Rich (1991) for Australia, Bertrandand Mullainathan (2004) for the Unites States, and Duguet and Petit (2005); Duguet, Leandri, L’Horty, and Petit (2010), Edo and Jacquemet (2013) for France.

10 However, it halves or reduces the number of observations even more compared to an experiment based on the usual correspondence test on a similar numberof job openings.

11 In addition, recruiters could log into Facebook using an assumed name with no mention of their position as recruiter.12 See Appendices B and C, for examples, of resumes and cover letters. Additional material is available on request.13 For empirical evidence that employers tend to hire the worker with the lowest unemployment duration, see Eriksson and Rooth (2014). Using Swedish data,

they show that a recent period of unemployment has a negative effect on applicant callback rates.14 During the experiment we checked regularly that our fictitious profiles were the sole results. After our experiment was concluded a profile named Thomas

Marvaux, unrelated to our experiment, showed up on a less popular social network site, Badoo.com and then disappeared. This profile was not present duringthe time of the experiment. Therefore, a recruiter could not have seen this profile in addition to our Facebook profile. Also, Facebook has made additionalchanges to the layout since the experiment took place, which is why the screenshots in this paper may not correspond to the current Facebook page layout.

15 The market shares of the web search engines in France in December 2012 were as follows: 90.1% for Google, 3.3% for Bing, and 1.5% for Yahoo (source:http://www.atinternet.com, retrieved June 2014).

16 See Blanchard & Landier (2002) and Lamy Social (2012) for more details on French employment protection rules.

MANANT ET AL. | 241

17 We distinguish the variable mean wage offer from the other variables in the model due to multicollinearity with the other job characteristics.18 The condition index of Model 2 is equal to 29.8. This is very close to the threshold indicating multicollinearity problems according to Belsley, Kuh, and

Welsch (1980). This multicollinearity issue is not surprising since the wage offered is related to the required education level, and to job characteristics such asrequired experience, industry, job position, and firm size among others. Once we exclude mean wage offer in Model 3, the condition index falls to 20.4.

19 More descriptive statistics on jobs and firms’ characteristics are available in Appendix 2 of Supporting Information materials.20 The gap between the callback rates of the two control applicants at around 20 days of applications is due to the fact that, as callbacks start, a small difference

in the number of calls from employers causes a large gap in the cumulative rate. This effect disappears after a month of applications, and the rate of returnsstabilizes thereafter.

21 For further information, see Yao R. 2013. “Improvements to Timeline,” http://newsroom.fb.com/News/584/Improvements‐to‐Timeline, retrieved June 2017;Milano D. 2012. “Facebook May Be Changing Your Timeline: Redesign Tests in Progress” http://abcnews.go.com/blogs/technology/2012/12/facebook‐may‐be‐changing‐your‐timeline‐redesign‐tests‐in‐progress, retrieved June 2017, and on other news, or bloggers’ websites that observed this change about mid‐December 2012 (retrieved June 2017): http://mashable.com/2012/12/20/facebook‐timeline‐change//#GoRYX74dYPqU and http://www.marismith.com/facebook‐single‐column‐timeline‐layout/.

22 The corresponding Fisher’s exact test is below the 10% significance level (Pr = 0.069). These model‐free evidence are confirmed using a probit model (seeAppendix 4 in Supporting Information materials). All the models—base model and models including control variables—show a significant and negativeimpact on callback rates of the Arabic applicant.

23 More descriptive statistics on jobs and firms’ characteristics are available in Appendix 3 of Supporting Information materials.24 Robustness checks provided in Appendix 5 of Supporting Information materials confirm that applicant’s type is not significant for this period.25 We thank an anonymous referee for her helpful comments in this discussion.

ORCID

Matthieu Manant http://orcid.org/0000-0003-2691-9368Serge Pajak http://orcid.org/0000-0002-4652-1368Nicolas Soulié http://orcid.org/0000-0002-5045-3054

REFERENCES

Acquisti, A., & Fong, C. M. (2016). An experiment in hiring discrimination via online social networks. SSRN Working Paper.Aeberhardt, R., Coudin, E., & Rathelot, R. (2010). Les écarts de taux d’emploi selon l’origine des parents: Comment varient‐ils avec l’âge et le

diplôme? In INSEE (Ed.), France, portrait social. Paris: INSEE Edition, 149–166.Ahmed, A. M., Andersson, L., & Hammarstedt, M. (2013). Are gay men and lesbians discriminated against in the hiring process? Southern

Economic Journal, 79, 565–585.Ahmed, A. M., & Hammarstedt, M. (2008). Discrimination in the rental housing market: A field experiment on the internet. Journal of Urban

Economics, 64, 362–372.Albrecht, J. W., & van Ours, J. C. (2006). Using employer hiring behavior to test the educational signaling hypothesis. The Scandinavian

Journal of Economics, 108, 361–372.Arrow, K. J. (1973). Higher education as a filter. Journal of Public Economics, 2, 193–216.Autor, D. H. (2001). Wiring the labor market. The Journal of Economic Perspectives, 15, 25–40.Bartoš, V., Bauer, M., Chytilová, J., & Matějka, F. (2016). Attention discrimination: Theory and field experiments with monitoring

information acquisition. The American Economic Review, 106, 1437–1475.Battu, H., & Zenou, Y. (2010). Oppositional identities and employment for ethnic minorities: Evidence from England. The Economic Journal,

120, F52–F71.Bayer, P., Ross, S. L., & Topa, G. (2008). Place of work and place of residence: Informal hiring networks and labor market outcomes. Journal

of Political Economy, 116, 1150–1196.Becker, G. S. (1957). The theory of discrimination. Chicago: University of Chicago Press.Belsley, D. A., Kuh, E., & Welsch, R. E. (1980). Regression Diagnostics: Identifying Influential Data and Sources of Collinearity. New York:

John Wiley and Sons.Bertrand, M., & Duflo, E. (2017). Field experiments on discrimination. In A. Banerjee & E. Duflo (Eds.), Handbook of field experiments, (Vol.

1, pp. 309–394). Elsevier.Bertrand, M., & Mullainathan, S. (2004). Are Emily and Greg more employable than Lakisha and Jamal? A field experiment on labor market

discrimination. The American Economic Review, 94, 991–1013.Bisin, A., Patacchini, E., Verdier, T., & Zenou, Y. (2008). Are Muslim immigrants different in terms of cultural integration? Journal of the

European Economic Association, 6, 445–456.Blanchard, O., & Landier, A. (2002). The perverse effects of partial labour market reform: Fixed‐term contracts in France. The Economic

Journal, 112, F214–F244.Booth, A., & Leigh, A. (2010). Do employers discriminate by gender? A field experiment in female‐dominated occupations. Economics Letters,

107, 236–238.

242 | MANANT ET AL.

Brown, C., & Gay, P. (1985). Racial discrimination: 17 years after the act (646–647). London: Policy Studies Institute.Bursell, M. (2007). What’s in a name? A field experiment test for the existence of ethnic discrimination in the hiring process. DiVA Working Paper.Cappellari, L., & Tatsiramos, K. (2015). With a little help from my friends? Quality of social networks, job finding and job match quality.

European Economic Review, 78, 55–75.Clark, L. A., & Roberts, S. J. (2010). Employers use of social networking sites: A socially irresponsible practice. Journal of Business Ethics, 95,

507–525.DeVaro, J. (2008). The labor market effects of employer recruitment choice. European Economic Review, 52, 283–314.Doleac, J. L., & Stein, L. C. (2013). The visible hand: Race and online market outcomes. The Economic Journal, 123, F469–F492.Drydakis, N. (2009). Sexual orientation discrimination in the labour market. Labour Economics, 16, 364–372.Duguet, E., Leandri, N., L’Horty, Y., & Petit, P. (2010). Are young French jobseekers of ethnic immigrant origin discriminated against? A

controlled experiment in the Paris area. Annals of Economics and Statistics, 99/100, 187–215.Duguet, E. & Petit, P. (2005). Hiring discrimination in the french financial sector: An econometric analysis on field experiment data. Annals

of Economics and Statistics, 78, 79–102.Edelman, B., Luca, M., & Svirsky, D. (2017). Racial discrimination in the sharing economy: Evidence from a field experiment. American

Economic Journal: Applied Economics, 9, 1–22.Edo, A., & Jacquemet, N. (2013). Discrimination à l’embauche selon l’origine et le genre: Défiance indifférenciée ou ciblée sur certains

groupes? Économie et Statistique, 464, 155–172.Edo, A., Jacquemet, N., & Yannelis, C. (2015). Language skills and homophilous hiring discrimination: Evidence from gender‐and racially‐

differentiated applications. Review of Economics of the Household, 1‐28.Eriksson, S., & Rooth, D.‐O. (2014). Do employers use unemployment as a sorting criterion when hiring? Evidence from a field experiment.

The American Economic Review, 104, 1014–1039.Firth, M. (1981). Racial discrimination in the british labor market. Industrial and Labor Relations Review, 34, 265–272.Fisman, R., & Luca, M. (2016). Fixing discrimination in online marketplaces. Harvard Business Review, 88‐95.Galarza, F. B., & Yamada, G. (2014). Labor market discrimination in Lima, Peru: Evidence from a field experiment.World Development, 58, 83–94.Granovetter, M. (1995). Getting a job: A study of contacts and careers. University of Chicago Press, Chicago.Holzer, H. J. (1987). Hiring procedures in the firm: Their economic determinants and outcomes. USA: National Bureau of Economic Research

Cambridge, Mass.Hubbuck, J., Britain, G., & Carter, S. (1980). Half a chance? A report on job discrimination against young blacks in Nottingham. London:

Commission for Racial Equality.Ioannides, Y. M., & Loury, L. D. (2004). Job information networks, neighborhood effects, and inequality. Journal of Economic Literature, 42,

1056–1093.Jacquemet, N., & Yannelis, C. (2012). Indiscriminate discrimination: A correspondence test for ethnic homophily in the Chicago labor

market. Labour Economics, 19, 824–832.Kramarz, F., & Skans, O. N. (2014). When strong ties are strong: Networks and youth labour market entry. The Review of Economic Studies,

81, 1164–1200.Kuhn, P. J. (2014). The internet as a labor market matchmaker. IZA World of Labor.Kuhn, P., & Mansour, H. (2014). Is internet job search still ineffective? The Economic Journal, 124, 1213–1233.Lambrecht, A., & Tucker, C. E. (2018). Algorithmic bias? An empirical study into apparent gender‐based discrimination in the display of

STEM career ads. Management Science, forthcoming.Lamy Social (2012). Lamy social: Code du travail - édition 2012. Paris: Editions Lamy.Montgomery, J. D. (1992). Job search and network composition: Implications of the strength‐of‐weak‐ties hypothesis. American Sociological

Review, 57, 586–596.Obukhova, E., & Lan, G. (2013). Do job seekers benefit from contacts? A direct test with contemporaneous searches.Management Science, 59,

2204–2216.Oreopoulos, P. (2009). Why do skilled immigrants struggle in the labor market? A field experiment with six thousand resumes, USA:

National Bureau of Economic Research Cambridge, Mass.Oreopoulos, P. (2011). Why do skilled immigrants struggle in the labor market? A field experiment with thirteen thousand resumes.

American Economic Journal: Economic Policy, 3, 148–171.Pellizzari, M. (2010). Do friends and relatives really help in getting a good job? ILR Review, 63, 494–510.Phelps, E. S. (1972). The statistical theory of racism and sexism. The American Economic Review, 62, 659–661.Pope, D. G., & Sydnor, J. R. (2011). What’s in a picture? Evidence of discrimination from prosper.com. Journal of Human Resources, 46, 53–92.Rees, A. (1966). Information networks in labor markets. The American Economic Review, 56, 559–566.Riach, P. A., & Rich, J. (1991). Testing for racial discrimination in the labour market. Cambridge Journal of Economics, 15, 239–256.Riach, P. A., & Rich, J. (2002). Field experiments of discrimination in the market place. The Economic Journal, 112, F480–F518.Spence, M. (1973). Job market signaling. The Quarterly Journal of Economics, 87, 355–374.Sterling, A. D. (2014). Friendships and search behavior in labor markets. Management Science, 60, 2341–2354.Sunstein, C., & Thaler, R. (2008). Nudge. Yale University Press, New Haven.Van Ommeren, J., & Wentink, D. (2012). The (hidden) cost of employer parking policies. International Economic Review, 53, 965–978.Weichselbaumer, D. (2003). Sexual orientation discrimination in hiring. Labour Economics, 10, 629–642.

MANANT ET AL. | 243

SUPPORTING INFORMATION

Additional supporting information may be found online in the Supporting Information section at the end of the article.

How to cite this article: Manant M, Pajak S, Soulié N. Can social media lead to labor market discrimination?Evidence from a field experiment. J Econ Manage Strat. 2019;28:225–246. https://doi.org/10.1111/jems.12291

APPENDIX A: JOB APPLICATIONS BY INDUSTRY AND BY LOCATION (MARCH–SEPTEMBER 2012)

Control applicant Test applicant

Industry (%)Accounting 10.4 11.7Association/union 3.1 3.5Transport 5.7 4.3Bank/insurance 2.6 4.8Construction 2.6 3.1Retail trade 5.2 6.5Wholesale trade 10.0 10.8Audit/consulting 5.6 6.9Culture/leisure 1.8 1.3Management 4.4 3.0Teaching/research 5.2 3.4Hotel/restaurant 6.1 4.3Real estate 3.0 4.3Telecom/computer 4.4 5.2Health/social 6.5 5.6Public organizations 1.7 2.2Advertising/communication 4.8 4.3Business services 7.4 7.4Personal services 3.0 2.2Industry/energy/waste 6.5 5.2

Location (%)Seine‐et‐Marne 4.4 3.0Yvelines 9.1 8.6Essonne 4.8 9.1Hauts‐de‐Seine 16.1 15.5Seine‐Saint‐Denis 10.0 12.1Val‐de‐Marne 12.2 8.2Val‐d’Oise 4.4 3.5Central districts of Paris 9.1 8.2North‐East districts of Paris 6.5 5.2North‐West districts of Paris 11.7 13.8South‐East districts of Paris 5.2* 1.3*Applicant’s district and contiguous districts in Paris 6.5 8.6

Observations 230 232

Note. *indicates a significant difference in proportion at the 5% threshold between the two applicants.

Paris areas definitions:

• Central districts of Paris: 1st, 2nd, 3rd, 4th, and 5th districts;• North‐East districts of Paris: 10th, 11th, 19th, and 20th districts;• North‐West districts of Paris: 8th, 9th, 17th, and 18th districts;• South‐East districts of Paris: 12th and 13th districts; and• Applicants’ district and contiguous districts: 15th, 14th, 16th, 17th, 6th, and 7th districts.

244 | MANANT ET AL.

APPENDIX B: EXAMPLE OF A SENT RESUME (TRANSLATED INTO ENGLISH)

MANANT ET AL. | 245

APPENDIX C: EXAMPLE OF A SENT COVER LETTER (TRANSLATEDINTO ENGLISH)

246 | MANANT ET AL.